1
Leveraging EHR to Discover Effective
Care Coordination Practice Patterns
Session 4, February 12, 2019
You Chen, Assistant Professor, Vanderbilt University Medical Center
2
You Chen, PhD
has no real or apparent conflicts of interest to report.
Conflict of Interest
3
• Learning objectives
• High healthcare cost in the U.S.
• Fragmented vs. coordinated care
• Team science in coordinated care
• A data revolution to team science
• Introduction of data in electronic health records
• Learn effective teams and patient medical needs from the data
• Limitations and challenges of using the data in team science
• Three case studies to illustrate the learning of effective teams
• Conclusions
Agenda
4
Participants will be able to:
1. Describe how to use network analysis along with electronic health
records to show coordination patterns between providers
2. Identify how to leverage electronic health records to measure
patient medical needs
3. Measure relationships of team models with health outcomes
Learning Objectives
5
U.S. has the highest healthcare
cost among developed countries
79.1
$9,237
Per Capita Health spending (2014)
Average Life Expectancy
Source: Institute for
Health Metrics and
Evaluation, World
Bank country
classifications
6
Health cost has been increasing
exponentially in the U.S.
0 500000 1000000 1500000 2000000 2500000 3000000
1960
1963
1966
1969
1972
1975
1978
1981
1984
1987
1990
1993
1996
1999
2002
2005
2008
2011
2014
Total National Health Expenditures (Millions)
Source: The National
Health Expenditure
Accounts (NHEA)
7
Why healthcare cost is so high in
the U.S.?
Fragmented care
8
An example of fragmented care
He saw each doctor independently and each doctor
looked in depth at the organ, did the latest tests and
prescribed the latest drugs and devices
He bought the best of
each commodity
Cardiologist
Neurologist
Gastroenterologist
9
Fragmented care brings …
lack of appropriate medication
use and adherence
duplicative use of services
increase healthcare
expenditures
Governments, health care
systems, and individuals
spend more and more on
healthcare, for less and less
value
10
Coordinated care
Consider the whole person
rather than each of the discrete
diseases/organs
A group of providers interact
with each other rather than
each of them acts individually
11
Coordinated VS. Fragmented
Care
Fragmented care
Coordinated care
Source: IOM (Institute of Medicine). The Healthcare Imperative: Lowering Costs and Improving Outcomes: Workshop
Series Summary. Washington, DC: The National Academies Press, 2010.
12
Team science in coordinated
care
All providers treating a patient should be communicating
and sharing information to ensure that everyone is acting
as a team to meet the patient’s medical needs.
13
Dynamic changes of patient
medical needs
• require corresponding team
adjustments
Seldom researches have been
done to investigate self-organizing
nature of care team
• how do teams respond to the
dynamic changes of patient
medical needs?
• how do such responses relate to
patient outcomes?
14
Linking provider interaction
network structures to success
The Bavelas-Levitt Experiment
Star and Y: fastest problem solvers
Clear leader at the center of communications
15
Transforming data into provider
interaction networks
Provider
Patient
16
Data warehouse at Vanderbilt
The Synthetic Derivative (SD) and the Research Derivative (RD)
Distribution of SD subject population
by age, race and gender
IBM general parallel file
system for the Netezza
warehousing appliance
17
EHR audit logs
U
1
U
2
U
3
U
4
A physician
requested a
lab test
A lab user
uploaded a
lab test result
Physician
office
received the
lab test result
A nurse
provided
counseling
service
Nov,9, 2018,
9:00am
Nov,9, 2018,
4:00pm
Nov,9, 2018,
4:20pm
Nov,10, 2018,
10:20am
A Patient’s EHR
18
• No change to record (view)
– Form view or chart review
• Change to record (document)
– Medication or lab orders
– Notes
– Clinical actions-assessment, treatment or procedure
– Communications with providers or patients
Audit logs-overarching categories
19
An example of audit logs
Patient ID
Provider ID
Access action
Access date time
Capture interactions of
providers to patient EHRs
Form viewing, clinical notes
writing, medication
ordering, vital sign
monitoring ….
These events are rarely
viewed outside of a health
care system
20
Diagnosis data
Patients’ clinical notes,
billing codes,
laboratories, medication
orders and discharge
summaries
21
Examples of Patient Diagnoses
Represented by ICD-9 Codes
ID of a patient’s EHR
ICD-9 codes
assigned to a patient
An encounter ID
22
Leveraging audit logs to learn
care teams
A naïve way to
transform a bipartite
graph to an interaction
network of providers
23
An example to illustrate the
calculation of the interaction
strength
Strength (u
1
-> u
2
) = 4/7
p
2
p
3
p
7
p
5
p
4
p
6
p
1
p
2
p
3
p
7
p
5
u
1
u
2
Strength (u
2
-> u
1
) = 4/4
24
Matrix decomposition to learn
provider networks from audit
logs
25
Principal component analysis
2
nd
Principal Component
Children’s Hospital
University Hospital
1
st
Principal Component
26
Social Network Metrics
In-degrees
Out-degrees
Betweenness Centrality
Closeness Centrality
27
Leveraging diagnosis data to
weight patient medical needs
p
1
p
2
p
3
p
5
p
4
p
6
d
1
d
2
d
1
d
2
d
3
d
3
Assignments of
diagnoses to patients
28
TF-IDF weights medical needs of
a patient
The weight of a
disease specific to
a patient
Patient population
having disease d
Number of times the
disease appearing in a
patient's EHRs
29
An example of TF-IDF
Patient 1
Patient 2
Patient 3
Problem
1
4
2
4
Problem 2
0
0
4
Problem 3
0
0
3
Problem 4
1
0
0
Patient 1
Patient 2
Patient 3
Problem
1
0.27
0.27
0.11
Problem 2
0
0
0.28
Problem 3
0
0
0.28
Problem 4
0.69
0
0
TF-IDF
Common problem
Specific problem
Different weights
on health problems
30
Latent dirichlet allocation (LDA)
to learn topics weighting patient
medical needs
Health problem
Weight to a patient
31
How to evaluate the learned
care teams and medical needs?
Unsupervised
Learning
• No labeled data
• No prior knowledge
32
• Survey
– Recruit experts with domain knowledge to assess plausibility
of learned patterns: care teams/patient medical needs
• Simulation
– Simulate care teams/patient medical needs
– Test performances of models on the simulated data
Online surveys or computer
simulations
33
REDCap: an online survey system
• Open a new project
• Design survey questions
• Ask experts to answer questions
• Record survey results
• Analyze survey results
34
• Statistical models: proportional-odds model, analysis of variance,
logistical regression
– Patient outcome: length of stay, unplanned readmission, …
– Outcome ~ α + β
1
×team factors + (β
2
× patient
demographics + β
3
× patient medical needs + β
4
× health
insurance programs + β
5
× admission months + … )
Effectiveness of care teams
35
Challenges of using EHR audit data
Sign a note Appointment
cancelled
Accessing system
Taking care of patients
36
Extracting access actions which
are only related to patient care
Measurements
and laboratory
tests
Observations and procedures
Medication management
37
Challenges of using EHR diagnosis data
Providers’ EHR utilization behaviors vary
Provider generated data
38
Using Standard terminology to
represent diagnosis data
39
• Inpatient setting
Where to start?
Three case studies
40
Case study 1
learn care teams and their
responsible medical needs
41
• 10 659 VUMC employees
• 17 947 inpatients
• 5 176 unique ICD-9 billing codes
• 831 721 unique interactions the providers committed to EHRs of
patients
• 74 192 assignments of diagnosis codes to patients
4 months of audit logs and
diagnoses from VUMC
42
Provider interaction network
structures at the level of VUMC unit
43
27 of 34 care teams were confirmed
44
Matches between care teams
and patient medical needs
Phenotypes
Care teams
45
Oncology care team
Operational area
Distance between operational areas
Bone Marrow
Related
Radiation
Oncology
Related
Hematology and
Myelosuppression
Cancer
Center
46
Phenotypes associated with
oncology care team
47
Case study 2
Learning provider interaction
networks in the neonatal
intensive care unit and
measuring their relationships
with length of stay
48
• 1 day before surgery until 30 days after surgery
• 70 – 2400 provider actions recorded per patient
• Discharged to home or remained at hospital at 30-day post-
surgery mark
• Age, weight, birth date, race, gender, surgery date
EHR data of 18 infants with
gastrostomy surgery in NICU
49
Patient-Level Provider Networks
50
Out-degree average was
significantly related to post-
operative length of stay
• Out-degree is a measure of
information dispersion
• Patients treated by providers who
disseminated patient-related
information to providers within the
network had shorter LOS
High out-degree Low out-degree
Post surgical LOS (days)
51
Case study 3
Trauma provider interaction
networks and their relationships
with length of stay
52
• 5,547 employees committed EHR access actions during 5,588
patient encounters
• 158,467 unique actions
• Confounding factors including a patient’s age, historical
service utilization, diagnoses, procedures, admission season
and insurance program
EHR data of 5,588 adult
inpatient episodes hospitalized
survivors of trauma
53
Inpatient setting
The left interaction network has the highest degree of collaboration
between care providers and it was related to the shortest length of
stay
Graph density: 0.27 0.17 0.18
0.61 days shorter in average
54
• Health data science just starts…
• We are at the very beginning of the research on care teams and
patient medical needs
• While data-driven methods can provide insight into care team and
patient medical need modeling, there are still many challenges
need to be solved
– EHR data quality, security and privacy
– Generalizability of models: multiple healthcare organizations
– Evaluation of the learned novel knowledge
– The gap between the learned knowledge and its application
in clinical practice
– Intra-coordination vs. inter-coordination
Conclusions
